Bug 94782

Now Qt's ImageDiff generate a grayscale image based on distance of actual and expected pngs: https://trac.webkit.org/browser/trunk/Tools/DumpRenderTree/qt/ImageDiff.cpp?rev=124324#L98 Zero distance is black, white-black distance is white on the diff image now. Unfortunately there are: - many failing tests on different distros (for example Ubuntu and Debian because of different libfreetype system package): minor font differences because of anti-aliasing - many failing tests on 32/64 bit systems: All pixels of a green rectangle are same green(128) on 64 bit, but the down margin line of this rectangle are lighter green(127) It causes minor pixel differences: (1,1,1) coloured line on the diff image, which is absolutely undetectable now. And in most cases NRWT reports this failures as "pixel hash failed (but diff passed)", because Qt's ImageDiff now ignores differences smaller than 0.01%. difference. 0.01% is 48 full pixel difference (black/white) now. (800x600 png) But a 800 pixel width (127/128, 0/0, 0/0) difference means only 1/256 * 800 = 3.125 px distance. That's why we get ~1800 "pixel hash failed (but diff passed)" failures on 32/64 bit builds. To detect and debug these failures, it would be great if - Qt's ImageDiff marked all different pixels as red (255,0,0) - same as Chromium's ImageDiff - Qt's ImageDiff marked same pixels darker as the original (r/2, g/2, b/2) - same as Chromium's ImageDiff - tolerance was based on the number of different pixels instead of the euclidean distance Proposed patch and examples are coming soon.